In this position paper, we argue that the classical evaluation on Natural Language Processing (NLP) tasks using annotated benchmarks is in trouble. The worst kind of data contamination happens when a Large Language Model (LLM) is trained on the test split of a benchmark, and then evaluated in the same benchmark. The extent of the problem is unknown, as it is not straightforward to measure. Contamination causes an overestimation of the performance of a contaminated model in a target benchmark and associated task with respect to their non-contaminated counterparts. The consequences can be very harmful, with wrong scientific conclusions being published while other correct ones are discarded. This position paper defines different levels of data contamination and argues for a community effort, including the development of automatic and semi-automatic measures to detect when data from a benchmark was exposed to a model, and suggestions for flagging papers with conclusions that are compromised by data contamination.
Translate-test is a popular technique to improve the performance of multilingual language models. This approach works by translating the input into English using an external machine translation system, and running inference over the translated input. However, these improvements can be attributed to the use of a separate translation system, which is typically trained on large amounts of parallel data not seen by the language model. In this work, we introduce a new approach called self-translate, which overcomes the need of an external translation system by leveraging the few-shot translation capabilities of multilingual language models. Experiments over 5 tasks show that self-translate consistently outperforms direct inference, demonstrating that language models are unable to leverage their full multilingual potential when prompted in non-English languages. Our code is available at https://github.com/juletx/self-translate.
Humans can learn to understand and process the distribution of space, and one of the initial tasks of Artificial Intelligence has been to show machines the relationships between space and the objects that appear in it. Humans naturally combine vision and textual information to acquire compositional and spatial relationships among objects, and when reading a text, we are able to mentally depict the spatial relationships that may appear in it. Thus, the visual differences between images depicting "a person sits and a dog stands" and "a person stands and a dog sits" are obvious for humans, but still not clear for automatic systems. In this project, we propose to evaluate grounded Neural Language models that can perform compositional and spatial reasoning. Neural Language models (LM) have shown impressive capabilities on many NLP tasks but, despite their success, they have been criticized for their lack of meaning. Vision-and-Language models (VLM), trained jointly on text and image data, have been offered as a response to such criticisms, but recent work has shown that these models struggle to ground spatial concepts properly. In the project, we evaluate state-of-the-art pre-trained and fine-tuned VLMs to understand their grounding level on compositional and spatial reasoning. We also propose a variety of methods to create synthetic datasets specially focused on compositional reasoning. We managed to accomplish all the objectives of this work. First, we improved the state-of-the-art in compositional reasoning. Next, we performed some zero-shot experiments on spatial reasoning. Finally, we explored three alternatives for synthetic dataset creation: text-to-image generation, image captioning and image retrieval. Code is released at https://github.com/juletx/spatial-reasoning and models are released at https://huggingface.co/juletxara.
Automatic Image Caption Generation model that uses a CNN to condition a LSTM based language model.